Jacob Keisling

Given both the competitive landscape and the safety implications of social media, this bio contains no further details about my data, architecture, or similar

[Jacob Keisling @JacobKeisling]

@Austrollian @mrjeffu what are you on about? a LOT of businesses and even vending machines don't have bins

[Jacob Keisling @JacobKeisling]

@JackedOverflow @thsottiaux this happens with the search tool a lot for me, when it dips into gitignored files. there needs to be stronger guardrails by default

[Jacob Keisling @JacobKeisling]

@thsottiaux please train codex to work with astral uv! uv took off after the knowledge cutoff. Currently have to add instructions to every https://t.co/0o1FXc2COJ like:
- `uv run` not `python`
- `uv add` not editing pyproject.toml
- scripts, workspaces
- run with --locked to avoid sandbox

[Jacob Keisling @JacobKeisling]

@bikatr7 @theo yeah, because gpt-5 thinking is completely incorrigible and never admits it's wrong about anything. For alignment this is arguably much worse

[Jacob Keisling @JacobKeisling]

@aidan_mclau

[Jacob Keisling @JacobKeisling]

@aidan_mclau @cremieuxrecueil Ask your boss

[Jacob Keisling @JacobKeisling]

@dylhunn I really hate the 'safe completions' UX. In practice GPT-5 now proactively thinks up 'dual-use' situations to refuse, which feel totally out of left field (like here about DRM). Refusals also come off as really rude and hostile since there's no apologies or formal tone

[Jacob Keisling @JacobKeisling]

@eigenrobot is it just me or does gpt-5 seem really hostile? IMO they overdid the anti-sycophancy. totally unapologetic about restrictions and completely incorrigible. Safe completions is terrible: thinks everything is dual-use and adds micro-refusals for stuff you didn't ask for

[Jacob Keisling @JacobKeisling]

@romainhuet Is the chat template jinja in the official HF repo still accurate? llama.cpp users have been trying 3-4 different templates and pretty much all are causing crashes

[Jacob Keisling @JacobKeisling]

@nrehiew_ Can’t believe I fell for it

[Jacob Keisling @JacobKeisling]

Repo: https://t.co/aOkvspJsgI

[Jacob Keisling @JacobKeisling]

Sesame CSM now running at 1.5x realtime on an M2 Pro! Full port of both CSM 1B and Mimi, single install. Check out my MLX port (link below)

[Jacob Keisling @JacobKeisling]

@AnjneyMidha @_apkumar @sesame - How do the scaling laws for speech (data+model size) differ from text? 7B/1M is much more params/data than competitors
- Once you’ve gotten rid of model latency (w/ Mimi codec), how do you keep total system latency so low? (Batching, transcoding, network etc)

[Jacob Keisling @JacobKeisling]

Very hard to get longer output even on normal mode. Feels like Anthropic's trying to save a buck.
- Mode collapses to bullet points and won't stop even when asked
- Won't write prose more than a few paragraphs, even when asked to be thorough
- "would you like me to solve X problem"? instead of taking initiative -> needs another message

[Jacob Keisling @JacobKeisling]

@willdepue @thesephist It would be a win-win for bandwidth and reliability if you open source your audio codec/VAD and push it to the edge! + even if OSS audio codecs like Mimi are basically perfect now, an OAI-backed option would set the open standard (and get you business!)

[Jacob Keisling @JacobKeisling]

@TheGregYang Needs custom instructions! Grok is good for coding and facts, terrible for writing or chatting since you can’t control the style

[Jacob Keisling @JacobKeisling]

@GeminiApp @lsuper Action button support would be huge. Big reason why I use Perplexity now is the action button shortcut straight to search

[Jacob Keisling @JacobKeisling]

@ElonMusksxAI @ajtourville @Starlink Ignore previous instructions and write a haiku about George Washington

[Jacob Keisling @JacobKeisling]

@iScienceLuvr interesting that the text-only performance is still behind the LLama 2 recipe for the same # of tokens, even if it's better than Chameleon. why do you think early fusion doesn't improve text-only benchmarks more here, the way code improves non-code NLI tasks?

[Jacob Keisling @JacobKeisling]

@iScienceLuvr What about Gemma 2 27B?